

########## Version 0.12-5 (April 2015) ##################

* S3 methods added to the NAMESPACE file

########## Version 0.12-4 (September 2014) ##################

* Several small changes to .cpp files as a result of CRAN warnings
* The old src/Makevars file was deleted

########## Version 0.12-3 (July 2013) #######################

* R version 3.0 changed the "eigen" function, so that if the "symmetric=TRUE" 
option is not given explicitely, *and if rownames and colnames are not the 
same* it is assumed that the matrix is *not* symmetric. This was causing 
complex eigenvalues to appear in the submatrices of the "trim.matrix" 
function (only the main matrix was given the "symmetric=TRUE" option.
The bug was reported by a user on 13.6.13.
* Subdirectory "vignettes" was created for vignette sources, as suggest post R 2.14.
* Lines wider than 90 characters were cut down in *.Rd files.

########## Version 0.12-2 (August 2012) #######################

* Bug correction in the eleaps function. In versions 0.11-1 through
0.12-1, for a few data conditions the processing of the variables 
provided in include argument was interfering with the checks of 
singularity. Version 0.12-2 fixes this problem.


########## Version 0.12-1 (July 2012) #######################

* patch for non-standard C++ code which gave an error mesage when compiled on Solaris platforms.


########## Version 0.12-0 (July 2012) #######################


* Changes were made to the code, so as to allow also the anneal, improve and 
genetic algorithms to keep running, when called with criteria "tau2", "xi2", 
"zeta2" or "ccr12", if a subset associated with a singular "mat" matrix is 
selected. Whereas until now this would break the execution of the search 
algorithms, as of now an artificially poor criterion value is associated with 
such subsets (to ensure that they are nor selected), but the algorithm proceeds. 
At the end, a warning is passed on to the user if some subsets associated with
ill-conditioned submatrices were indeed found and dropped in the course of the
searches. 

* The memory management of the Fortran and C++ code was improved, making 
the anneal, improve, genetic and eleaps funtions more robust when handling 
larger (up to a few hundred original variables) problems.

* An additional search sequence was made available to the eleaps function, making
it more adaptable to characteristics of the data. In particular, when the original 
"mat" matrix is numerically singular and optimality is not likely to be proved within 
the specified timelimit, eleaps now uses a diferent sequencing that tries to ensure 
that the potentially more interesting subsets are evaluated first.

* The numerical procedures for detecting and handling ill-conditioned matrices were 
improved.

* Two bugs were corrected in eleaps function: (i) The argument "timelimit" was not working
correctly with large (several thousand seconds) values. (ii) When called with non-NULL 
include and exclude arguments, searches according to the RV criterion were not working
correctly. 

* The default value for the "popsize" argument in the genetic function was changed from
100 to max(100,2*ncol(mat)), trying to ensure that for larger problems reasonable genetic
diversity is available by default.

* The output returned by genetic function, when stopped because the initial population 
did not guarantee enough genetic diversity, was changed. Previously, in this case  
genetic returned NA, and now it returns the best solutions found in the search before 
it was aborted.

* The output given by searches that were not able to find numericaly reliable solutions 
for all cardinalities sought was simpliflied and restricted to the cardinalities where 
reliable subsets were found.

* Some warning messages are now displayed more selectively and only in the presence of 
non-expected problems. The warning message about ill-conditioned matrices is only shown 
when it is not known in advance that the input matrix "mat" was singular (in particular, 
this message is not displayed when it is known that "mat" was originated from a problem 
with more original variables than observations). Likewise, the warning regarding restricted 
searchs is now displayed only if ill-conditioned subsets were indeed dropped in the course 
of the search.

* The lda method for the pre-processing function ldaHmat was dropped. This method worked 
incorrectly if between creating an MASS lda object and calling the ldaHmat function, the
x or grouping objects were changed. 
 

########## Version 0.11-3 (December 2011) #######################

* More changes in the C++ and R code to try and avoid remaining errors in 
the Solaris compilation and Notes in five other compilations.

* Bug correction in eleaps function: the change made in July 2011 to allow 
for "partial" searches when the mat matrix is singular was also conducting 
partial searches when the mat matrix was OK. 


########## Version 0.11-2 (October 2011) #######################

* Minor changes in the C++ and R code to try and avoid errors in the Solaris compilation, a warning in Fedora compilation and Notes in all other compilations.

* "using namespace std" added to several *.cpp files.


########## Version 0.11-1 (July 2011) #######################

* The "leaps" function was renamed "eleaps", to distinguish it from 
other "leaps" functions in R packages. The former "leaps" function is 
deprecated, but a "leaps" wrapper for the 
eleaps function was created in order to ensure backward compatibility.

* A Sweave vignette was created in the new inst/doc subdirectory.

* Changes were made to the code, so as to allow the eleaps search algorithm
to keep running when called with criteria tau2, xi2, zeta2 or ccr12,
if a subset associated with a singular mat matrix is selected. 
Whereas until now this would break the execution of the search 
algorithms, as of now an artificially poor criterion value is associated 
with such subsets (to ensure that they are nor selected), but the algorithm 
proceeds. At the end, a warning is passed on to the user. This change has 
not yet been introduced in the anneal, improve and genetic algorithms.

* A "data" directory was created. A "farm" data set and its help file were
created.

* A new attribute "FisherI" created, to indicate that an object is of 
class glm and that the Wald criterion should be the default criterion.

* A Subversion repository was created for the subselect project


########## Version 0.10.1 (December 2009) ###################

* A new criterion for variable selection in a generalized linear
model context: the Wald criterion. Files wald.R (in /R) and wald.Rd
(in /man) were added. 

* All search algorithm files (anneal.R, genetic.R, improve.R and
leaps.R) were changed to cater for the Wald criterion.

* All algorithm man files were changed to account for the Wald
criterion. 

* The default criterion has been changed to "default" and is set to
"RM" if r=0 and to "TAU2" if r>0.

* Changes were made to the validation.R file, to cater for the Wald
criterion. 

* For compatibility with the output of other search functions, 
"drop=FALSE" was added to the subsetting of the output of the "genetic" 
function, when "kabort" is involved. 

* Changes were made to ensure that the C++ code followed the standards. 
Problems had arisen with gcc 4.4 and the SunStudio compilers.

########## Version 0.9.9993 (January 2009) ###########

Changes in the C++ code which broke the package.
("#include <cstdlib>" added in sscma.cpp)

########## Version 0.9.9992 (February 2007) ###########

Changes in the Fortran code which broke the package on Linux 64-bit machines
(e.g.: conversions between types were smoothed out).


########## Version 0.9.9991 (January 2007) ###########

Changes in the C code which could break the package on Linux 64-bit machines
(e.g.: use of "long" data type was changed to "int").
 
########## Version 0.9.999 (October 2006) ############

1) Changes in the C code required by the transition to R 2.4 (and the
use of new C compiler).

2) The leaps.R, lmHmat.default and glhHmat.default functions have
suffered minor changes for greater numerical stability.

3) The "qr" criterion used when checking for positive definiteness of
the input matrix in the validmat.R function was removed for coherence
with the use of the "tolval" argument. 


########## Version 0.9.99 (June 2006) ############

1) Three new functions have been added to "pre-process" the data for
submission to the search algorithms using any of the four new
"multivariate linear hypothesis" criteria: function \code{lmHmat}
creates the input for a linear model (linear regression), function
\code{ldaHmat} does the same for a linear discriminant analysis and
\code{glhHmat} does likewise for the general linear hypothesis.

2) The \code{call} item in the output list of the \code{leaps}
function (which had been mistakenly dropped in version 0.9-9) was
re-added. 


########## Version 0.9.9 (May 2006) ############

Some major changes, as well as some minor ones.

1) Four new criteria have been added: "CCR1_2", "TAU_2", "XI_2" and
"ZETA_2". All are criteria for subset selection in the context of the
multivariate linear model. The four criteria were added as options for
the "criterion" argument in the subset selection algorithms (anneal,
genetic, improve and leaps). 

2) Four new functions were created to compute the value of the new
criteria when given a total matrix, a model effects matrix (H) with
its expected rank (r) and a subset of variables. These new functions
are called "ccr12.coef", "tau2.coef", "xi2.coef" and "zeta2.coef", and join
the similar functions "gcd.coef", "rm.coef" and "rv.coef" which already
existed. 

3) Two new auxiliary functions were created: "validmat" and
"validnovcrit", which include validation checks for the symmetry and
positive definiteness of the total matrix and for the suitability of
the H and r arguments of the new criteria, respectively. These are not
to be called directly by the user.

4) The validation part of the old and new functions has been further
modularized, with calls to the new validation functions mentioned above.

5) A new argument "tolsym", with default value
1000*.Machine$double.eps, has been added to the four search functions
(anneal, genetic, improve and leaps). It is used in the symmetry
checks of the total and effects (H) matrices. When the supplied
matrices have corresponding elements differing by more than this value,
they are rejected. If corresponding elements differ, but by less than this
value, the supplied matrix is replaced by its symmetric part and a
warning is provided.

6) Man files were written for the new functions, and the help files
for the search functions were improved, with examples using the new
criteria. 



########## Version 0.9.1 (June 2005) #############

1) The default value of argument "tolval" in functions "trim.matrix",
"leaps", "anneal", "genetic" and "improve" has been changed to
10*.Machine$double.eps (previously .Machine$double.eps).

2) Correction of minor mistakes in the "leaps" help page and in
function "validation" (which is not in the Namespace) - changes made
to the "leaps" function between versions 0.8 and 0.9 had not been updated
in this function and help page.

########## Version 0.9 (March 2005) #############

1) Namespace added.

2) A new function "trim.matrix" was introduced, to deal with the issue
of multicollinearities in the input data.

3) The four search functions ("leaps", "anneal", "genetic" and
"improve") now perform a test for multicollinearity, which consists
in comparing the ratio of the smallest to the largest eigenvalue of
the input covariance/correlation 
matrix with the new argument "tolval" (.Machine$double.eps, by
default). If this 
ratio is smaller than "tolval", the function exits with a
message to the user requiring that the input matrix be cleared of its
ill-conditionment (perhaps through the use of the new function
"trim.matrix") before resubmitting it to the search function. The use
of the search functions can be forced, by lowering the value of
"tolval" (negative values of "tolval" are not allowed), but with
risks of numerical inaccuracy.  

4) The previous maximum number of variables allowed for the search
functions (300) was eliminated. The code now runs for any number of
variables.  A test for input data with more than
400 variables was introduced in the search functions, and when that
condition tests positive, the function stops execution and passes a
message to the user, who may choose to run the function anyways, by
setting a new function argument ("force") to TRUE. The code will now
run with any number of variables, but may crash the R session due to
memory problems. 

5) Non-standard C++ code in the routines was changed and code which
caused memory errors was also modified. The new code passes the
"valgrind" memory checks.

6) Some Fortran code which did not comply with Fortran 77 
standards was changed (e.g.: symbols such as ">" changed to
".GT."; comments using "*" in the first column; etc.). Some
non-standard code was kept (use of  
variable array sizes in subroutine declaratives; use of do-loops and
do-while-loops). 

7) The default behaviour when a non-square matrix is passed as input data
to the search functions has been changed: instead of assuming that
covariance matrix of the input data matrix was wanted, the new option
is to compute the input data matrix's *correlation* matrix.

8) The default values for "pcindices" whas been changed in 
the "leaps" function and is now "first_k" as in the "anneal",
"genetic" and "improve" functions. This change is possible due to 
changes in the "leaps" source code.

########## Version 0.8 (May 2004) ##############

1) Output lists for the four search functions (leaps, anneal,
genetic and improve) now have a fifth object: "$call", which gives the
match.call() of the command that produced the output.

2) The C++ routine to compute the eigendecomposition for the "GCD"
criterion in the "leaps" function has been improved. A more exact
routine based on the QR decomposition is now used (see file
"qldiagon.cpp" in the /src subdirectory).

3) Validation of input for the four search functions has now been
transfered to separate modules ("validation.R", "validannimp.R" and
"validgenetic.R"), which are mentioned in
"/man/subselect-internal.Rd". Some minor changes to these validation
checks were made. 

4) The default values for "pcindices" were changed in 
the "leaps" function (where it is now NULL) and in the "anneal",
"genetic" and "improve" functions (where it is now "first_k"). This
seeks to highlight the behaviour of the functions when the GCD
criterion (which uses "pcindices") is invoked with the default value
of pcindices: "leaps" explicitely requires a set of pcindices, whereas
the other three search functions will, by default, compare, for each
cardinality "k" that has been requested, the k-variable subsets with
the first k PCs. 


########## Version 0.7.1 (March 2004) ##############

1) Two lines defining constants "TRUE" and "FALSE" in the C++ file
src/matrixb.h, which gave errors when compiling in Mac OS X, were
changed. The same change was made in file SR_vsda.h, and similar
changes with the constants "EPSLON" were made in files lagmat.cpp,
gaussjel.cpp, gausstrg.cpp, matfact.cpp, pdiagon.cpp and
simgaussjell.cpp. Constants "MIN" and "MAX" in files SR_vsda.h,
sr_sscma.cpp and sr_wrkf1.cpp were renamed MINIMZ and MAXIMZ,
respectively.

2) Warnings in the R code and help files were introduced, to caution
against the difference in default behaviour of the "leaps" function -
when compared to the search algorithm functions "anneal", "genetic"
and "improve" - in case the "GCD" criterion is requested. In the latter
functions, GCD values compare (by default) the subspaces spanned by
k-variable subsets  and by the first k PCs. In the leaps function, GCD
values, by default, compare the subspaces spanned by k-variable
subsets and by the first *kmin* PCs (i.e., the cardinalities of the
subsets of variables and PCs are not equal, except for the first
cardinality requested). This option is directly related to the nature of the
leaps algorithm.


########## Version 0.7 (March 2004) ##############

1) A major new function has been included: "leaps". Leaps implements
Duarte Silva's adaptation of Furnival and Wilson's Leaps and Bounds
Algorithm for variable selection in Regression Analysis. The bulk of
computations are carried out by C++ routines.

2) New checks for symmetry and positive-definiteness of the input
matrix were added.

3) The references in the help files were updated. 



########## Version 0.4.1 (February 2004) #########

1) Changed file gcd.Rd, which had old default option for the argument
pcindices. This generated a codoc mismatch.



##########  Version 0.4 (April 2003) ##############

1) Version 1.7 of R includes LAPACK. The configure scripts to check
   for the presence of LAPACK in the system have been deleted.

2) Argument "initialsol" in functions "anneal" and "improve", as well
   as argument "initialpop" in function "genetic" are now NULL by
   default. The validations and initializations have been changed as a result.

3) Argument "pcindices" in function "gcd.coef" is initially NULL and
   will be set to 1:kmax unless the user has specified a set of PC indices.

4) Validation tests in functions "gcd.coef", "rm.coef" and "rv.coef"
   to check whether user-specified variable (and or PC) indices are
   integers.

5) Validation tests in functions "anneal", "improve" and "genetic" to
   test whether inital solutions or population are specified as arrays
   of integers.

6) A bug was fixed in functions "anneal", "genetic" and "improve" in
   those cases when argument "pcindices" is defined by the user. A
   logical comparison between a vector "pcindices" and another vector
   was being treated as if a global value for equality were returned.


##########  Version 0.3 (October 2002) ##############

1) Changes to the configure scripts necessary for configuration in the
   MacOSX.

2) Functions anneal, genetic and improve now include checks for
   covariance/correlation matrices not of full rank. If non-full rank
   matrices are given, the functions exit with a warning.



##########  Version 0.2  (June 2002) #################

The following changes were made in Version 0.2 of R package subselect:

1) A bug was corrected in function anneal: when more than one solution
   (nsol > 1) was requested, the initial temperature for subsequent
   solutions was not being updated.

2) The Fortran code was changed to use R's default random number
   generator, introducing a call to R as described in pages 48 and
   50-51 of the "Writing R Extensions" manual, version 1.5. As a
   consequence, the LAPACK routine SLARUV is no longer used and the
   "iseed" argument of functions "anneal", "genetic" and "improve" has
   been replaced by the logical argument "setseed".

3) Initial solutions (variable subsets) may now be fed by the user to
   each of functions "anneal" (argument "initialsol"), "genetic"
   (argument "initialpop") and "improve" (argument "initialsol").

4) The Fortran code was changed to pass messages directly to the R GUI
   (which did not occur in the Windows version), as described in page 
   50 of the "Writing R Extensions" manual, version 1.5.

5) The "indices" argument of functions "gcd.coef", "rm.coef" and
   "rv.coef" can now be 2-d or 3-d arrays. In this way, the "$subsets"
   and "$bestsets" output options of functions "anneal", "genetic" and
   "improve" can be passed directly to functions "rm.coef", "rv.coef"
   and "gcd.coef", allowing easy computation of values of any
   criteria for subsets produced by the search algorithms using 
   a different criterion.

6) The "printfile" option of functions "anneal", "genetic" and
   "improve" has been dropped, as printing to files is best done using
   the R environment.

7) Function "anneal" has a new argument, "coolfreq", which controls
   the frequency with which the simulated annealing temperature is
   cooled (used to be hard-coded as every 20 iterations in the Fortran
   code of version 0.1).

8) Function "genetic" has a new argument "mutprob", which controls the
   probability of each offspring in the genetic algorithm undergoing a
   mutation (i.e., being fed to the restricted local improvement
   algorithm), if the "mutate" option is set to TRUE (previously *all*
   children were mutated, which considerably slowed down run times).

9) Function "full.k.search" has been dropped from the
   package. It was computationally inefficient and not in the general
   spirit of the package. For computationally efficient full-search
   algorithms, see program SSCMA by Pedro Duarte Silva of the
   Universidade Catolica do Porto, available at
   http://porto.ucp.pt/psilva

10) The help files were updated accordingly. Details of the "rv.coef"
   and "rm.coef" functions were also introduced.
